ICLR 2021 Papers

Neural networks (NNs) whose subnetworks implement reusable functions are expected to offer numerous advantages, including compositionality through efficient recombination of functional building blocks, interpretability, preventing catastrophic interference, etc. Understanding if and how NNs are modular could provide insights into how to improve them. Current inspection methods, however, fail to link modules to their functionality. In this paper, we present a novel method based on learning binary weight masks to identify individual weights and subnets responsible for specific functions. Using this powerful tool, we contribute an extensive study of emerging modularity in NNs that covers several standard architectures and datasets. We demonstrate how common NNs fail to reuse submodules and offer new insights into the related issue of systematic generalization on language tasks.

Despite —or maybe because of— their astonishing capacity to fit data, neural networks are believed to have difficulties extrapolating beyond training data distribution. This work shows that, for extrapolations based on finite transformation groups, a model’s inability to extrapolate is unrelated to its capacity. Rather, the shortcoming is inherited from a learning hypothesis: Examples not explicitly observed with infinitely many training examples have underspecified outcomes in the learner’s model. In order to endow neural networks with the ability to extrapolate over group transformations, we introduce a learning framework counterfactually-guided by the learning hypothesis that any group invariance to (known) transformation groups is mandatory even without evidence, unless the learner deems it inconsistent with the training data. Unlike existing invariance-driven methods for (counterfactual) extrapolations, this framework allows extrapolations from a single environment. Finally, we introduce sequence and image extrapolation tasks that validate our framework and showcase the shortcomings of traditional approaches.

Normalization techniques have proved to be a crucial ingredient of successful training in a traditional supervised learning regime. However, in the zero-shot learning (ZSL) world, these ideas have received only marginal attention. This work studies normalization in ZSL scenario from both theoretical and practical perspectives. First, we give a theoretical explanation to two popular tricks used in zero-shot learning: normalize+scale and attributes normalization and show that they help training by preserving variance during a forward pass. Next, we demonstrate that they are insufficient to normalize a deep ZSL model and propose Class Normalization (CN): a normalization scheme, which alleviates this issue both provably and in practice. Third, we show that ZSL models typically have more irregular loss surface compared to traditional classifiers and that the proposed method partially remedies this problem. Then, we test our approach on 4 standard ZSL datasets and outperform sophisticated modern SotA with a simple MLP optimized without any bells and whistles and having ~50 times faster training speed. Finally, we generalize ZSL to a broader problem — continual ZSL, and introduce some principled metrics and rigorous baselines for this new setup. The source code is available at https://github.com/universome/class-norm.

Differentially private SGD (DP-SGD) is one of the most popular methods for solving differentially private empirical risk minimization (ERM). Due to its noisy perturbation on each gradient update, the error rate of DP-SGD scales with the ambient dimension $p$, the number of parameters in the model. Such dependence can be problematic for over-parameterized models where $p \gg n$, the number of training samples. Existing lower bounds on private ERM show that such dependence on $p$ is inevitable in the worst case. In this paper, we circumvent the dependence on the ambient dimension by leveraging a low-dimensional structure of gradient space in deep networks---that is, the stochastic gradients for deep nets usually stay in a low dimensional subspace in the training process. We propose Projected DP-SGD that performs noise reduction by projecting the noisy gradients to a low-dimensional subspace, which is given by the top gradient eigenspace on a small public dataset. We provide a general sample complexity analysis on the public dataset for the gradient subspace identification problem and demonstrate that under certain low-dimensional assumptions the public sample complexity only grows logarithmically in $p$. Finally, we provide a theoretical analysis and empirical evaluations to show that our method can substantially improve the accuracy of DP-SGD in the high privacy regime (corresponding to low privacy loss $\epsilon$).

The privacy leakage of the model about the training data can be bounded in the differential privacy mechanism. However, for meaningful privacy parameters, a differentially private model degrades the utility drastically when the model comprises a large number of trainable parameters. In this paper, we propose an algorithm \emph{Gradient Embedding Perturbation (GEP)} towards training differentially private deep models with decent accuracy. Specifically, in each gradient descent step, GEP first projects individual private gradient into a non-sensitive anchor subspace, producing a low-dimensional gradient embedding and a small-norm residual gradient. Then, GEP perturbs the low-dimensional embedding and the residual gradient separately according to the privacy budget. Such a decomposition permits a small perturbation variance, which greatly helps to break the dimensional barrier of private learning. With GEP, we achieve decent accuracy with low computational cost and modest privacy guarantee for deep models. Especially, with privacy bound $\epsilon=8$, we achieve $74.9\%$ test accuracy on CIFAR10 and $95.1\%$ test accuracy on SVHN, significantly improving over existing results.

In this paper, we explore connections between interpretable machine learning and learning theory through the lens of local approximation explanations. First, we tackle the traditional problem of performance generalization and bound the test-time predictive accuracy of a model using a notion of how locally explainable it is. Second, we explore the novel problem of explanation generalization which is an important concern for a growing class of finite sample-based local approximation explanations. Finally, we validate our theoretical results empirically and show that they reflect what can be seen in practice.

Interactive Fiction (IF) games provide a useful testbed for language-based reinforcement learning agents, posing significant challenges of natural language understanding, commonsense reasoning, and non-myopic planning in the combinatorial search space. Agents based on standard planning algorithms struggle to play IF games due to the massive search space of language actions. Thus, language-grounded planning is a key ability of such agents, since inferring the consequence of language action based on semantic understanding can drastically improve search. In this paper, we introduce Monte-Carlo planning with Language Action Value Estimates (MC-LAVE) that combines a Monte-Carlo tree search with language-driven exploration. MC-LAVE invests more search effort into semantically promising language actions using locally optimistic language value estimates, yielding a significant reduction in the effective search space of language actions. We then present a reinforcement learning approach via MC-LAVE, which alternates between MC-LAVE planning and supervised learning of the self-generated language actions. In the experiments, we demonstrate that our method achieves new high scores in various IF games.

Recent empirical and theoretical studies have shown that many learning algorithms -- from linear regression to neural networks -- can have test performance that is non-monotonic in quantities such the sample size and model size. This striking phenomenon, often referred to as "double descent", has raised questions of if we need to re-think our current understanding of generalization. In this work, we study whether the double-descent phenomenon can be avoided by using optimal regularization. Theoretically, we prove that for certain linear regression models with isotropic data distribution, optimally-tuned $\ell_2$ regularization achieves monotonic test performance as we grow either the sample size or the model size. We also demonstrate empirically that optimally-tuned $\ell_2$ regularization can mitigate double descent for more general models, including neural networks. Our results suggest that it may also be informative to study the test risk scalings of various algorithms in the context of appropriately tuned regularization.

In the era of causal revolution, identifying the causal effect of an exposure on the outcome of interest is an important problem in many areas, such as epidemics, medicine, genetics, and economics. Under a general causal graph, the exposure may have a direct effect on the outcome and also an indirect effect regulated by a set of mediators. An analysis of causal effects that interprets the causal mechanism contributed through mediators is hence challenging but on demand. To the best of our knowledge, there are no feasible algorithms that give an exact decomposition of the indirect effect on the level of individual mediators, due to common interaction among mediators in the complex graph. In this paper, we establish a new statistical framework to comprehensively characterize causal effects with multiple mediators, namely, ANalysis Of Causal Effects (ANOCE), with a newly introduced definition of the mediator effect, under the linear structure equation model. We further propose a constrained causal structure learning method by incorporating a novel identification constraint that specifies the temporal causal relationship of variables. The proposed algorithm is applied to investigate the causal effects of 2020 Hubei lockdowns on reducing the spread of the coronavirus in Chinese major cities out of Hubei.

Recently proposed consistency-based Semi-Supervised Learning (SSL) methods such as the Pi-model, temporal ensembling, the mean teacher, or the virtual adversarial training, achieve the state of the art results in several SSL tasks. These methods can typically reach performances that are comparable to their fully supervised counterparts while using only a fraction of labelled examples. Despite these methodological advances, the understanding of these methods is still relatively limited. To make progress, we analyse (variations of) the Pi-model in settings where analytically tractable results can be obtained. We establish links with Manifold Tangent Classifiers and demonstrate that the quality of the perturbations is key to obtaining reasonable SSL performances. Furthermore, we propose a simple extension of the Hidden Manifold Model that naturally incorporates data-augmentation schemes and offers a tractable framework for understanding SSL methods.

Learning 3D geometry directly from raw data, such as point clouds, triangle soups, or unoriented meshes is still a challenging task that feeds many downstream computer vision and graphics applications. In this paper, we introduce SALD: a method for learning implicit neural representations of shapes directly from raw data. We generalize sign agnostic learning (SAL) to include derivatives: given an unsigned distance function to the input raw data, we advocate a novel sign agnostic regression loss, incorporating both pointwise values and gradients of the unsigned distance function. Optimizing this loss leads to a signed implicit function solution, the zero level set of which is a high quality and valid manifold approximation to the input 3D data. The motivation behind SALD is that incorporating derivatives in a regression loss leads to a lower sample complexity, and consequently better fitting. In addition, we provide empirical evidence, as well as theoretical motivation in 2D that SAL enjoys a minimal surface property, favoring minimal area solutions. More importantly, we are able to show that this property still holds for SALD, i.e., with derivatives included. We demonstrate the efficacy of SALD for shape space learning on two challenging datasets: ShapeNet that contains inconsistent orientation and non-manifold meshes, and D-Faust that contains raw 3D scans (triangle soups). On both these datasets, we present state-of-the-art results.

A challenging open question in deep learning is how to handle tabular data. Unlike domains such as image and natural language processing, where deep architectures prevail, there is still no widely accepted neural architecture that dominates tabular data. As a step toward bridging this gap, we present Net-DNF a novel generic architecture whose inductive bias elicits models whose structure corresponds to logical Boolean formulas in disjunctive normal form (DNF) over affine soft-threshold decision terms. Net-DNFs also promote localized decisions that are taken over small subsets of the features. We present an extensive experiments showing that Net-DNFs significantly and consistently outperform fully connected networks over tabular data. With relatively few hyperparameters, Net-DNFs open the door to practical end-to-end handling of tabular data using neural networks. We present ablation studies, which justify the design choices of Net-DNF including the inductive bias elements, namely, Boolean formulation, locality, and feature selection.

We develop an algorithm to train individually fair learning-to-rank (LTR) models. The proposed approach ensures items from minority groups appear alongside similar items from majority groups. This notion of fair ranking is based on the definition of individual fairness from supervised learning and is more nuanced than prior fair LTR approaches that simply ensure the ranking model provides underrepresented items with a basic level of exposure. The crux of our method is an optimal transport-based regularizer that enforces individual fairness and an efficient algorithm for optimizing the regularizer. We show that our approach leads to certifiably individually fair LTR models and demonstrate the efficacy of our method on ranking tasks subject to demographic biases.

Training Deep Neural Networks (DNN) with high efficiency can be difficult to achieve with native floating-point representations and commercially available hardware. Specialized arithmetic with custom acceleration offers perhaps the most promising alternative. Ongoing research is trending towards narrow floating-point representations, called minifloats, that pack more operations for a given silicon area and consume less power. In this paper, we introduce Block Minifloat (BM), a new spectrum of minifloat formats capable of training DNNs end-to-end with only 4-8 bit weight, activation and gradient tensors. While standard floating-point representations have two degrees of freedom, via the exponent and mantissa, BM exposes the exponent bias as an additional field for optimization. Crucially, this enables training with fewer exponent bits, yielding dense integer-like hardware for fused multiply-add (FMA) operations. For ResNet trained on ImageNet, 6-bit BM achieves almost no degradation in floating-point accuracy with FMA units that are $4.1\times(23.9\times)$ smaller and consume $2.3\times(16.1\times)$ less energy than FP8 (FP32). Furthermore, our 8-bit BM format matches floating-point accuracy while delivering a higher computational density and faster expected training times.

Though convolutional neural networks (CNNs) have demonstrated remarkable ability in learning discriminative features, they often generalize poorly to unseen domains. Domain generalization aims to address this problem by learning from a set of source domains a model that is generalizable to any unseen domain. In this paper, a novel approach is proposed based on probabilistically mixing instance-level feature statistics of training samples across source domains. Our method, termed MixStyle, is motivated by the observation that visual domain is closely related to image style (e.g., photo vs.~sketch images). Such style information is captured by the bottom layers of a CNN where our proposed style-mixing takes place. Mixing styles of training instances results in novel domains being synthesized implicitly, which increase the domain diversity of the source domains, and hence the generalizability of the trained model. MixStyle fits into mini-batch training perfectly and is extremely easy to implement. The effectiveness of MixStyle is demonstrated on a wide range of tasks including category classification, instance retrieval and reinforcement learning.

Greedy-GQ is a value-based reinforcement learning (RL) algorithm for optimal control. Recently, the finite-time analysis of Greedy-GQ has been developed under linear function approximation and Markovian sampling, and the algorithm is shown to achieve an $\epsilon$-stationary point with a sample complexity in the order of $\mathcal{O}(\epsilon^{-3})$. Such a high sample complexity is due to the large variance induced by the Markovian samples. In this paper, we propose a variance-reduced Greedy-GQ (VR-Greedy-GQ) algorithm for off-policy optimal control. In particular, the algorithm applies the SVRG-based variance reduction scheme to reduce the stochastic variance of the two time-scale updates. We study the finite-time convergence of VR-Greedy-GQ under linear function approximation and Markovian sampling and show that the algorithm achieves a much smaller bias and variance error than the original Greedy-GQ. In particular, we prove that VR-Greedy-GQ achieves an improved sample complexity that is in the order of $\mathcal{O}(\epsilon^{-2})$. We further compare the performance of VR-Greedy-GQ with that of Greedy-GQ in various RL experiments to corroborate our theoretical findings.

Interpretation of Deep Neural Networks (DNNs) training as an optimal control problem with nonlinear dynamical systems has received considerable attention recently, yet the algorithmic development remains relatively limited. In this work, we make an attempt along this line by reformulating the training procedure from the trajectory optimization perspective. We first show that most widely-used algorithms for training DNNs can be linked to the Differential Dynamic Programming (DDP), a celebrated second-order method rooted in the Approximate Dynamic Programming. In this vein, we propose a new class of optimizer, DDP Neural Optimizer (DDPNOpt), for training feedforward and convolution networks. DDPNOpt features layer-wise feedback policies which improve convergence and reduce sensitivity to hyper-parameter over existing methods. It outperforms other optimal-control inspired training methods in both convergence and complexity, and is competitive against state-of-the-art first and second order methods. We also observe DDPNOpt has surprising benefit in preventing gradient vanishing. Our work opens up new avenues for principled algorithmic design built upon the optimal control theory.

Concept-based explanation approach is a popular model interpertability tool because it expresses the reasons for a model's predictions in terms of concepts that are meaningful for the domain experts. In this work, we study the problem of the concepts being correlated with confounding information in the features. We propose a new causal prior graph for modeling the impacts of unobserved variables and a method to remove the impact of confounding information and noise using a two-stage regression technique borrowed from the instrumental variable literature. We also model the completeness of the concepts set and show that our debiasing method works when the concepts are not complete. Our synthetic and real-world experiments demonstrate the success of our method in removing biases and improving the ranking of the concepts in terms of their contribution to the explanation of the predictions.

Image and video synthesis are closely related areas aiming at generating content from noise. While rapid progress has been demonstrated in improving image-based models to handle large resolutions, high-quality renderings, and wide variations in image content, achieving comparable video generation results remains problematic. We present a framework that leverages contemporary image generators to render high-resolution videos. We frame the video synthesis problem as discovering a trajectory in the latent space of a pre-trained and fixed image generator. Not only does such a framework render high-resolution videos, but it also is an order of magnitude more computationally efficient. We introduce a motion generator that discovers the desired trajectory, in which content and motion are disentangled. With such a representation, our framework allows for a broad range of applications, including content and motion manipulation. Furthermore, we introduce a new task, which we call cross-domain video synthesis, in which the image and motion generators are trained on disjoint datasets belonging to different domains. This allows for generating moving objects for which the desired video data is not available. Extensive experiments on various datasets demonstrate the advantages of our methods over existing video generation techniques. Code will be released at https://github.com/snap-research/MoCoGAN-HD.